A federal learning backdoor defense method based on gradient screening and weight adjustment

Abstract

This invention discloses a backdoor defense method for federated learning based on gradient filtering and weight adjustment, relating to the field of federated learning security. The method first identifies and eliminates potential anomalous updates through gradient adaptive sparsity, cluster analysis, and anomaly detection to reduce redundant information and lower the risk of malicious attacks. Then, during the aggregation process, the weights of each client are adaptively adjusted based on factors such as gradient similarity, optimizing the model aggregation process and ensuring that the contributions of benign clients are effectively utilized while mitigating the impact of malicious clients. Compared with existing defense algorithms, this invention can effectively resist various backdoor attacks without significantly sacrificing model accuracy. It exhibits superior defense performance, robustness, and low computational complexity in different data environments and can be widely applied to various federated learning systems.

Images